Quality Externalities on Platforms:

The Case of Airbnb∗

Sonia Jaffe† Peter Coles‡ Steven Levitt§ Igor Popov¶

November 22, 2017

Click here for the most up-to-date version

Preliminary and IncompletePlease do not cite or circulate without authors’ permission.

Abstract

In order to screen out low-quality sellers or incentivize high quality, platforms need

a good measure of seller quality. Having high-quality sellers is particularly important if

buyers on a platform have limited information about the sellers; buyers’ learning means

that the quality of a seller in any given transaction affects whether or how frequently

that buyer returns to the platform. A seller’s influence on the number of times her

buyers return to the platform is an externality that the seller exerts on the platform’s

other sellers; we propose using this externality to measure seller quality. Using data

from Airbnb, a peer-to-peer accommodation platform, we calculate each listing’s guest

return propensity (GRP), defined as the average number of subsequent bookings a list-

ing’s guests complete, controlling for guest and trip characteristics. There is substantial

variation in GRP across listings and its correlation with a listing’s average rating is

only 0.05. Using an instrumental variable analysis to account for unobservable guest

characteristics, we find that our measure of GRP has a causal effect on returns: a one

standard deviation increase in GRP causes guests to take an additional 0.34 trips (a

17% increase). We discuss how platforms can increase overall seller surplus by directing

buyers towards higher quality sellers, either by using Pigouvian subsidies for quality

or by prioritizing high-quality sellers in their search algorithms.

∗The opinions and views expressed herein are those of the authors, not Airbnb. The authors thank Kevin

Murphy for helpful comments and Ryan Parks and Anya Marchenko for research assistance.†University of Chicago. Corresponding author: spj@uchicago.edu‡Airbnb§University of Chicago¶Airbnb

1

Platforms live or die by their ability to match users and align incentives. Online mar-

ketplaces such as Lyft, eBay, and Airbnb do not directly control production of their goods

or services, relying instead on market design decisions to manage their marketplaces. These

online platforms, along with more traditional ‘platforms’ such as newspapers and farmers’

markets, have inspired a robust economics literature on platform externalities, which focuses

almost exclusively on quantity-based externalities (e.g., eBay buyers benefit when new sellers

join the market and are potentially harmed by additional buyers). Less academic attention

has been paid to externalities that arise from network quality.

In contrast to the academic literature, online marketplaces invest heavily in managing

and incentivizing high quality for the goods and services provided through their platform.

In doing so, they face a key challenge: quality is notoriously difficult to measure. Platforms

can often measure purchase propensity. User-generated ratings vary in their informativeness

(whether the ratings actually differ across agents) and in their effectiveness, (whether users

base their decisions on the ratings). But even at their best, there is no clear mapping from

these metrics to the value a seller contributes to the platform.

In this paper, we explore the ‘quality externality’ that platform participants exert on

other users on the same side of the market and propose that platforms should measure

quality based on this externality. A marketplace participant that provides a good (bad)

experience makes it more (less) likely that the agent on the other side of the transaction

returns to the platform. These future transactions help other participants on the same side

of the market, but are not internalized by each agent.1 A platform can estimate the effect

of each seller’s behavior on the propensity of their buyers to continue interacting with the

platform. We call this metric the guest return propensity (GRP).

We begin by presenting a theoretical framework that motivates the use of GRP and guides

the empirical analysis. On the buyers’ side, users update their expectations about average

platform quality with each purchase. We extend the basic learning model to incorporate

learning about the relationship between price and quality. The key insight is that this

buyer-side learning can generate a wedge between individual seller incentives and those of

the platform or sellers as a group. It is likely that a buyer’s subsequent purchases on the

platform are from different sellers, so the initial seller does not internalize that benefit.

From the platform’s point of view, this setup mirrors a standard case for a Pigouvian tax,

highlighting the potential efficiency gains from aligning social and private incentives across

agents.

1The externality seems particularly important in a platform such as Groupon, where the business modelis based on buyers trying new sellers. If bad experiences cause buyers to only purchase Groupons for sellersthat they already frequent, then Groupons are less profitable for sellers.

2

We then present the empirical context. We describe our data from Airbnb, which coordi-

nates millions of trips in 191 countries. We observe rich information about search and travel

patterns on the platform over time. Using bookings data, we construct a GRP metric for

each accommodation listed on Airbnb. We find significant variation in GRP across listings,

even controlling for guest and trip characteristics. Very highly reviewed listings differ in their

guests’ propensity to rebook. In fact, GRP is only weakly correlated with guests’ rating of

that listing or with the number of guests the listing has.2 Moreover, the GRP measure is

persistent, predicting rebooking rates out of sample.

Our learning model predicts that GRP will have a stronger effect on the beliefs of new

users, who have weaker priors regarding platform quality. Consistent with this, we find

that GRP has a larger effect on the probability of return for inexperienced guests than

for experienced ones. However, GRP’s effect on total future guest spending is larger for

experienced guests. This difference reflects the fact that experienced guests are more likely

to be frequent travelers so, though the effect is smaller, it influences more potential future

trips and therefore more future spending on the platform. Thus, from the platforms point

of view, experienced guests are attractive targets for high-GRP listings, even if their beliefs

are less affected by GRP.

Using granular search data, we design an instrumental variable strategy to show that

the effect of a listing’s GRP on guests’ subsequent trips is causal rather than just reflecting

unobserved guest characteristics. Frequent bookings, changes in availability, and ongoing

search experiments often result in two users visiting the site on the same day entering the

same search terms being shown different inventory. The size of our sample allow us to

leverage these quickly changing choice sets for identification. We instrument for the GRP of

the booked listing with the average GRP across the first page of search results, using only

variation generated by multiple guests searching for the same market for the same travel

day, with comparable trip lead time.

We find that high-GRP listings cause a significant increase in guests’ future platform

utilization. In our instrumental-variable specification, (exogenously) booking a listing with

a one standard deviation higher GRP leads a guest to take 0.34 additional future trips.

This effect is three times larger than the comparable effect of a guest booking a listing

with a one standard deviation higher rating. In contrast to the effect on future travel,

displaying high-GRP listings in a guest’s search results does not lead to many more immediate

bookings. A one standard deviation increase in the average GRP displayed only raises

2The lack of correlation with ratings may seem surprising, but makes sense if we think that ratings arebased on things guests think are specific to a given listing, whereas whether a guest returns is not effectedby things they think are listing-specific, but precisely by the things they think generalize to all listings. Thelow correlation with the number of guests is consistent with the idea that GRP is not observable ex-ante.

3

purchase probabilities by 0.68% (0.00038 percentage points). We take this as evidence that

GRP reflects the quality of the actual experience and not necessarily some feature of the

listing that users observe while searching.

We end by discussing the implications of our results for Airbnb and how they might

generalize to other platforms. Airbnb could use our measure of listing quality to raise overall

seller surplus by removing low quality hosts from the platform, or by directly trying to

help low-quality hosts improve. Alternatively, it could incentivize higher listing quality and

redirect guests to higher-quality listings either by implementing a Pigouvian tax on low

quality or by showing guests higher quality listings in their search results.

To our knowledge, Nosko and Tadelis (2015) were the first to highlight the role of quality-

driven own-side platform externalities. They argue that reputational externalities arise when

buyers update their beliefs about the quality of all sellers on the platform from individual

interactions. In the context of the eBay platform, they propose a quality metric that takes

into account the propensity of buyers to leave feedback. They then use whether a buyer

returns to eBay to validate this measure of quality. We build upon their work in two key

ways. First, we provide a formal theoretical treatment of buyer learning and the implications

for a platform’s surplus. Second, we show that reputational externalities can be directly

converted into a useful quality metric. In this way, we offer a new, complementary quality

measure that is directly informed by reputational externalities.

More broadly, we contribute to the growing literature exploring two-sided market design

and platform externalities (e.g., Rochet and Tirole, 2003, 2006; Evans, 2003; Armstrong,

2006; Hagiu, 2007; Rysman, 2009; Weyl, 2010; Bardey et al., 2009) This body of work studies

platform design and pricing problems and platform competition in a setting where users exert

cross-side membership or usage externalities. We explore the existence and implications of

quality externalities in the same type of setting.

In focusing on users’ learning about platform quality from interactions with individ-

ual agents, we build upon the collective reputations and reputational commons literatures.

Shapiro (1982) models consumer learning and firm quality investments in equilibrium. Tak-

ing these concepts to an empirical setting, Landon and Smith (1998) estimate the effects

of individual and group reputational effects in the market for Bordeaux wines. King et al.

(2002) note that if individual firm actions can affect group reputation, a special case of the

commons problem exists (see also, Winfree and McCluskey, 2005; Barnett, 2007). We apply

these principles to the platform setting to understand how intermediaries can mitigate these

commons problems.

Finally, this paper relates to the literature exploring quality and ratings metrics used by

online marketplaces (e.g., Dellarocas, 2003; Cabral and Hortacsu, 2010; Mayzlin et al., 2014;

4

Fradkin et al., 2015; Luca, 2017). Our contribution to this literature is the introduction of a

new metric that is platform-independent and grounded in the theory of quality investment

and externalities.

The remainder of the paper is organized as follows. Section 1 outlines the theory, dis-

cussing both consumer learning and the externality it generates. Section 2 describes the

Airbnb context and data we use in the analysis. Here, we show how we construct the GRP

metric for each listing. Section 3 presents the empirical results, both the reduced-form anal-

ysis and the instrumental-variable approach using the search data. In Section 4, we discuss

potential strategies the platform could use to raise the average GRP of the listings that

guests visit. Section 5 concludes.

1 TheoryWe start with a basic model of consumers’ behavior and learning. We then consider the

future value of information and the possibility that consumers observe a price or other signal

prior to purchase. We then look at the platform objective function and how it compares to

the payoffs of individual sellers.

1.1 Individual ChoicesThe central aspects to our model of user choice are not specific to platforms: there is a

good whose quality is not known and varies across units; an individual may have multiple

opportunities to consume the good and each time she does she updates her prior on the

distribution of quality.

Consumers, indexed by i, first arrive in the market with a prior over the distribution of

the good’s quality, f0. Each period, with probability πi, a consumer has an opportunity to

potentially consume the good; a consumer with the opportunity to consume gets a draw,

ηt ∼ G, which is her value for a good of zero quality (a normalization).3 In addition,

consumers get utility u(q) if they purchase the good and its quality is q ≷ 0, for some

concave function u(·).Myopic Individuals

Start with the case of consumers who are myopic and base their purchase decision only

on the current period’s utility; they will purchase the good if

ηt + E[u(qt)] > 0.

This implies a threshold that depends on the beliefs that period, where the consumer pur-

chases if and only if η > η(·), where η is a function of individual’s beliefs about the dis-

3Alternatively, ηt can be thought of as the the observable component of quality of the available seller,where observable and unobservable quality are uncorrelated.

5

tribution of quality. If a consumer purchases the good, she observes a quality draw qt and

updates her prior accordingly.

For simplicity, assume that the prior is normal.4 In the most basic model, consumers may

know (or think they know) the variance of quality in the market, but be uncertain about

the mean. In this context, their prior can be described by three parameters: the variance

in quality across sellers, 1/γ, the individual’s belief about the average quality, µ0, and the

confidence they have in that belief, τ0.

The first time a consumer has an opportunity to purchase the good, the cutoff will be

η = −E[u(q1)|q1 ∼ N

(µ0,

1

τ0+

1

γ

)].

If she chooses to purchase, she will get a draw q1 and update her prior to

µ1 =τµ0 + γq1τ0 + γ

τ1 = τ0 + γ.

The less variance there is seller quality, the more weight the individual puts on a quality

draw.

Lemma 1. The more confidence consumers have in their beliefs (higher τ) or the more

variation there is across sellers in quality (lower γ), the less an individual’s beliefs will be

influenced by a given quality draw.

Lemma 1 suggests that quality externalities may be smaller on a platform like Etsy5 –

where the artistic nature of the products may lead consumers to think that sellers differ

a lot – then a platform like Uber where all the “sellers” are offering a fairly homogeneous

product – “a ride” – so consumers may have less reason to expect big differences across

sellers. Similarly, if consumers have stronger priors because a platform has been around

longer or is more widely known, a single purchase will affect their beliefs less.

In addition to not knowing the mean level of quality across sellers, individuals may be

uncertain about its variance. In this case, their prior consists of four components: a belief

about the mean and variance of the quality distribution – µ0, σ0 – and the precision of those

beliefs – τ0 and d0, respectively. Each time a consumer observes a quality draw, qk, she

4Without the assumption that u(·) is concave, the assumption of a normal prior would just be a normal-ization; since quality has no inherent unit, it could be re-scaled to be normally distributed. The assumptionsof concavity and normality together are substantive, but they allow us to get interesting comparative staticswith closed-form results.

5Etsy is a platform for selling primarily handmade or vintage items and craft supplies.

6

updates her beliefs to a posterior

µk =µk−1τk−1 + qkτk−1 + 1

τk = τk−1 + 1

σk =dk−1σk−1 + τk−1

τk−1+1(qk − µk−1)2

dk−1 + 1

dk = dk + 1

where k indexes only those periods in which the individual purchased the good. An individ-

ual’s purchase threshold depends only on the mean and variance, η(fk) = η(µk, σk). Later

posteriors are less effected by the quality draw.

Lemma 2. When a consumer has more past quality draws, that is k > k′, then

1. The mean of the consumer’s posterior is less effected by an additional draw: That is∂µk∂q

<∂µk′∂q

.

2. If the prior mean and variance are the same, and the consumer’s threshold cutoff η is

decreasing in the quality draw,6 then the cutoff will be less sensitive to the quality draw

when the consumer has more past quality draws. That is if µk = µk′ and σk = σk′ then

0 >∂η(µk, σk)

∂q>∂η(µk′ , σk′)

∂q.

When q < µ, increasing the quality both increases the mean of the posterior and decreases

its variance; these two changes both push η down, but less so when a consumer is more

experienced. If q > µ, then increasing the quality increases the variance of the posterior,

which has the opposite effect on η as increasing the mean. We cannot sign the derivative∂η

∂qand so the sign of the derivative is uncertain and sensitive (smaller absolute value of the

derivative) to a quality draw when she has more previous quality draws.

Though a given individual’s beliefs will be more effected by observed quality when she

has fewer past draws, individuals who have many past purchases are not randomly selected

from the population; they are more likely to be individuals who get lots of opportunities to

purchase – those with a high πi. The effect on an individual’s probability of purchase – as

distinct from beliefs – is

πig(η)∂η

∂q.

6While this is the case in general, at very high levels of quality it is the increase in posterior variancefrom raising quality can push the threshold up more than the increase in the posterior mean pushes it down.

7

If we only observe one period’s decision, then the expectation of an individual’s type, E[πi]

is higher if they purchase than if they do not.7 In general, we may expect more experienced

travelers to be positively selected on π. Therefore, the effect of quality on their behavior

could be larger than for inexperienced consumers, even if the effect on their beliefs is smaller.

Forward-looking consumers

Forward looking consumers have the same process for learning about quality, but a dif-

ferent decision rule. With a prior fk, the utility of purchase relative to not purchase is

W (η|fk) = η + E[u(q)] + δ(Eq

[W (fk+1 (fk, q))

]− W (fk)

),

where W (fk) = πi1−(1−πi)δEη[max{0,W (η|fk)}] and fk+1(fk, q) is the prior updated after ob-

serving q. This is similar to a single-armed bandit problem, though the (known) value of

the outside option changes from period-to-period.8 Non-myopic consumers will have a lower

threshold η for purchasing because they recognize that purchasing the good this period pro-

vides information value for future periods. Overtime, as the precision of fk increases, that

information value will decrease and forward looking consumers will behave more like myopic

consumers.

Prices

In many markets goods vary not only in quality, but also in price. Allowing goods

to vary by price means individuals may learn about the distribution of price and, more

importantly, its correlation with quality.9 In a contrived setting where a price-quality pair

(q, p) is randomly drawn, updating can follow a simple rule. If the beliefs at the beginning of

period t are means µt = (µqt , µpt ), variance-covariance σ2

t =

(σ2q,t σ2

qp,t

σ2qp,t σ2

p,t

), and confidence

7If one observes a sequence of purchase choices, it is actually possible that the individual with morepurchases is likelier to have a lower πi. If the true mean of the distribution is above the prior, then oncean individual purchases once, we expect her to purchase more, because on average the posterior beliefs willindicate better quality than the prior. If an individual purchases in period 1 and then does not purchasefor many periods, then we think the individual must have a low πi (and the first period was fluke) becauseotherwise she would have returned in a subsequent period; whereas if an individual does not purchase in thefirst period, her lack of subsequent purchasing can be explained by the low prior, so she may still have afairly high πi.

8This implies that unlike the simple single-armed bandit, a user may chose not to purchase in one period,but then chose to purchase in subsequent periods.

9It may also mean that some purchases may be more valuable if they transact at a higher price, whichwe discuss in Section 1.2.

8

τt, and a consumer draws (q, p)t+1, the consumer’s posterior beliefs will be

µt+1 =τtµt + (q, y)t+1

τt + 1, τt+1 = τt + 1,

σ2t+1 =

τtτt + 1

(σ2t +

1

τt + 1((q, y)t+1 − µt) ((q, y)t+1 − µt)T

).

As one would expect, the believed correlation increases whenever (q − µqt+1)(p− µpt+1) > 0.

If, more realistically, a consumer chooses from among a set of goods, after observing their

prices, then the consumer is not learning about the distribution of price, but only about the

distribution of quality and its correlation with price. In this case, there is no conjugate prior

for the beliefs about the correlation, so rather than look at the updating process, we focus

on the effect of a consumer’s belief about the correlation between prices and quality on her

decision of whether to purchase and what price-level of the good to choose.

For notational convenience, let v(µ, σ) = E [u(q)|q ∼ N(µ, σ)] be the expected value given

the beliefs about the mean and variance of quality. Because u(·) is concave, v(·) is increasing

in the mean and decreasing in the variance and ∂2v∂µ2

< 0, ∂2v∂µ∂σ

> 0. For a given correlation

ρ and price p, the distribution of quality is q ∼ N(µqt + (pj − µpt )ρ, (1− ρ2)σ2

q )). Increasing

the correlation has two effects for the consumer. First, it either increases or decreases the

expected quality, depending on whether p is above or below the mean price. Second, it

decreases the variance of the distribution of quality conditional on price, which is always

good.

If the consumer’s overall utility is v(µ, σ)−p (quasilinear in price), then her optimal price

satisfies

∂v

∂µρ = 1.

When the correlation changes, we get

∂p∗

∂ρ=

∂v∂µ

− ∂2v∂µ2

ρ2+

(p− µp)ρ

− 2σ

∂2v∂µ∂σ

− ∂2v∂µ2

(1)

The first effect is a direct effect of higher correlation pushing for a higher p because an

increase in p brings a larger corresponding increase in (expected) q. However, if the optimal

p was already above the mean, that the increase in expected quality (for a fixed p) decreases

the marginal return to expected quality, so if the correlation is positive, the second term

is negative. Lastly, higher correlation implies a lower variance in q conditional on p; this

decrease in variance means choosing a lower price involves less left-tail risk, which also pushes

for a lower optimal price. In our empirical analysis we look at whether the effect of quality

9

differs by the price of the listing.

A similar analysis applies for other characteristics that the consumer may observe prior to

purchase, and can therefore base their decision on. Consumers could also update their priors

on the correlation between quality and ratings or any other ex-ante observable characteristic.

The quasi-linearity we used for prices, would be less reasonable, but there would still be

competing effects. If rating and quality are more highly correlated, it means a consumer

gets more quality for a given increase in rating, which pushes for a higher rating; it also

means that a given high rating predicts higher quality, so if there are decreasing returns to

quality, a consumer may choose a lower quality.

1.2 Externality

Consumers learning about quality means that higher quality this period results in more

purchases in later periods. What is special to the platform context (since learning about

quality also applies to a single seller) is that an individual seller will only receive a fraction of

the returning consumer’s later purchases, and does not care about the additional purchases

that other sellers receive. Therefore, there is a quality externality : the private benefit to a

seller of having higher quality is lower than the social benefit.10

If there is no variation in price so both the sellers and the platform are just trying to

maximize purchases, then the platform’s payoff from consumer i is

W = πi

∞∑t=1

δtE[G(η(q1 . . . qk(t)|µ0, σ0, τ0, d0

))],

where δ is the discount factor and k(t) is the number of purchases a consumer has made at

the start of time t. Each seller j gets some share, dj of sales and a corresponding share dj

of the payoff

ωj = djπi

∞∑t=1

δtE[G(η(q1 . . . qk(t)|µ0, σ0, τ0, d0

))]= djW.

A seller’s quality affects the consumer’s cutoff in period t directly, but can also have an

indirect effect via the other purchase decision the consumer makes in the intervening periods.

If a good seller in period 1 causes the consumer to purchase (with higher probability) in

period 2, then in period 3, the consumer’s prior is based on the quality observed in period 1

and 2 instead of just the quality observed in period 1. Nevertheless, the seller’s incentive to

10This differs somewhat from the “reputation commons” problem because even if buyers can differentiateamong sellers, they still update their beliefs about the distribution of quality among other sellers. Badquality will make them less likely to purchase even if they know a specific seller was not the one who hadlow quality in the previous period. As long as the sellers are not available every period (or at some of thelocations the consumer goes to purchase), then good quality will also benefit other sellers.

10

improve quality is also dj of the social value.

Proposition 1. The value to the seller of improving quality in period 1 is

∂ωj∂qj

= djπi

∞∑t=1

δt∂

∂qj

(E[G(ηt

)])dj

where ηt

= ηt

(q1 . . . qk(t)|µ0, σ0, τ0, d0

). The value to the seller is the social value times the

seller’s share

∂ωj∂qj

= dj∂W

∂qj.

The share dj enters the seller’s value twice – once for the probability that the consumer

buys from that seller in period 1 and once for the probability that the buyer returns to that

seller. The first instance of dj also enters the platform’s marginal value of quality, since the

seller’s quality is only relevant if the buyer purchases from that seller; the latter instance of

dj does not enter the platform’s value because they care whether the guest returns, even if

it is to a different seller. The larger the seller’s share, the less misaligned her incentives are,

but the larger the effect on welfare of a decrease in her quality.

If either δ is small or quality effects the cutoff primarily in the periods directly subsequent

to purchase, then the value to the platform of a seller’s quality (relative to zero quality) for

a consumer who purchases in period 1 is just

W (qj)−W (0) = δE[G(η

2(qj|·))−G(η(0|·))

].

At the other extreme, if δ → 1, the value to the platform of a seller’s quality is

W (qj)−W (0) =∑t

E[G(η

t(qj|·))−G(η

t(0|·))

].

We use this measure of quality in our empirical analysis. We measure a seller’s quality as

the number of purchases a consumer makes after purchasing from them, controlling for the

number of purchases predicted for that consumer.11

Prices

If goods also vary in price then the platform may care about revenue instead of the

quantity of transactions. A seller’s quality may affect not just whether a consumer purchases,

but which product-price she chooses. However, that can all be summarized in the seller’s

11For computational convenience, we do not discount future trips by how far in the future they are. Allthe trips are within a few years and three quarters of those who return do so within 6 months, so while δ = 1is an approximation, a high δ is reasonable.

11

effect on total spending. In the empirical analysis we look at the effect of a listing’s quality

on subsequent spending by a guest on the platform:

W (qj)−W (0) = E

[∑t

pj(t)|q1 = qj

]− E

[∑t

pj(t)|q1 = 0

].

where if the consumer does not purchase in period t then j(t) = 0 and p0 = 0.

2 DataTo get a sense of the potential magnitude and variance in quality externalities, we use

data from Airbnb, a global peer-to-peer accommodations platform. Airbnb matches travelers

looking for accommodations to hosts who offer their home, apartment, extra bedroom, or

other accommodation to guests. Hosts can list their space, set their pricing and availability,

and accept bookings. Guests search for where they wish to stay and rate the home and

experience after traveling.

The Airbnb platform is large, providing ample data; there are over four million active

Airbnb listings in 191 countries. Figure 3 shows the geographical distribution of Airbnb

listings. Airbnb data also allow us to follow agents in the network, and their platform

behavior, over time. This information is essential for understanding return propensities and

the effect that quality may have on decision making. The other advantage of this context

is that accommodation quality is heterogeneous and not well-observed ex-ante. The market

has the opportunity to incorporate quality information into its marketplace design, making

this a relevant setting for platform policy with respect to quality.

While limiting to a subset of locations might give us a more homogeneous set of trips,

we want to be able condition on a guest’s travel history at Airbnb, so we use trips in all

locations. Similarly, though we have many trips, the number of listings with a large number

of trips is smaller, so in order to have more precision in estimating listing effects, we use all

guests.

Throughout our analysis, we use ‘trip’ to refer to a check-in by a guest at a listing; guests

might book multiple such ‘trips’ for a single episode of travel. The summary statistics are in

Table 1. Markets (which can include multiple cities) have had an average of 33.5K trips and

7.5K listings that have hosted a trip. The average rating of listings as of January 1, 2016

was 4.55. For users who have taken a trip, the average number of trips taken is 2.4; about

55% are female and the average age is 36. For trips since 2011, the average cost is $96 per

night. On average trips are booked just over a month in advance.

We are particularly interested in whether guests return to Airbnb after taking a trip.

Figure 2 shows the fraction of guests that return by trip number, for trips that occurred at

12

Figure 1: Log density of listings per square mileNote: These maps show the density of Airbnb listings in the United States and Europe. Listings per square

mile is calculated for each cell of .1◦ longitude by .1◦ latitude, using the approximation of 53 miles per degree

latitude.

Table 1: Summary of users, listings, trips, and markets

Statistic Mean St. Dev. Min Median MaxMarketsListings 1,900 7,137 1 303.5 165,151Trips 33,578 137,987 1 3,769 2,998,067ListingsRating 4.55 0.58 1.00 4.71 5.00Trips 18 35 1 5 1,524GuestsAge 35.82 12.30 18 32 99Female (0/1) 0.55 0.50 0 1 1Trips 2.36 2.93 1 1 475TripsNum of Guests 2.53 1.70 1 2 104Days in advancebooked 33.88 43.01 0 18 1,097

Nightly Price 96 94 0 73 32,216

Note: This table summarizes data about the listings, guests, trips and markets. All variables refer only to

listings and guests with at least 1 trip. The rating is the average rating as of January 1, 2016.

least 400 days before the end of the data. About a third of first time guests return within

400 days; that fraction climbs to over 95% for guests with at least 25 trips.

For part of our analysis, we use data on searches made on Airbnb’s website. There are

many more searches than bookings or trips. We drop any searches the platform thinks were

13

●

●

●

●

●●

●●

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

0.00

0.25

0.50

0.75

1.00

0 5 10 15 20 25

Trip Number

Ret

urn

Rat

e

Figure 2: Fraction of guests that return to Airbnb marketsNote: The graph shows the fraction of guests that take an additional trip within 400 days after their first,

second, etc trip. The dot for 25 also includes all trips after the 25th. It is based only on trips that happened

at least 400 days prior to the end of the data.

likely performed by a bot and restrict the data to the first search a potential guest made for

a given market and check-in day. Of these ‘first searches’ we drop any that had a minimum

or maximum price. We also drop searches for the same user and market with a check-in date

within two days, since these are in some sense not the user’s first search. We are left with

approximately 30 million searches per month in 2016. To keep the analysis manageable, we

use only 1 month of searches when looking at all searches and use 6 months when looking

only at searches that resulted in a booking.

Table 2 summarizes the search variables. The first section of Table 2a is characteristics

of the search or searcher, the second describes the search results and the third describes

the subsequent outcomes we link to the searches. About 78% of searchers are not people

the site recognizes as having previously booked a trip.12 On average they search for trips

77 days before the searched-for check-in date. Unsurprisingly, the results on average have

a substantially higher price than the trips described in Table 1 (people prefer lower-priced

listings); the ratings are also slightly higher. The construction of the guest return propensity

(GRP) variable is explained below. We only have quality and rating measures for a subset

of listings, so we weight the measured average quality (rating) by the share of listings that

have a quality (rating) measure, using the mean quality (rating) for the remaining listings.

12The site either does not recognize the user or the user is linked (by logging in or by the site recognizingtheir device) to an account with zero previous trips.

14

If, within 6 days of that initial search, a user booked a trip for the market searched for,

with a check-in date within 1 day of the date searched for, we say that search ‘resulted in’

that booking. If there are multiple such bookings, we take the first one. As we see in the

last section of Table 2, about 3% of searches result in a booking. Much of the analysis is, of

necessity, limited to the searches where we can identify the user id of the searcher; users are

identified if they log in or if the website recognizes their device. For those users, we match

their searches to the set of trips they booked and took subsequently. Searchers we can link

take an average of 1.8 trips following the search, spending $646 dollars. For searches by

identified users, 5.7% result in bookings.13

Table 2b summarizes some variables that are only available for the booked searches. The

average GRP and rating booked are slightly higher than the average shown. The users whose

searches we link to a booking spend an average of $360 on the linked trip and take an average

of 2.0 trips, spending $583, after the linked trip.

2.1 Quality

In addition to the raw data, we construct a quality measure for each listing.14 As discussed

in Section 1.2, we measure host quality as the guest return propensity : the number of trips a

guest takes after staying with that host (controlling for guest characteristics). For each trip

taken prior to 2016 by guest i at listing l in period t, we calculate the number of trips the

guest booked and took subsequently to get the outcome variable, yilt. We then regress

yilt = α0 + αl + α1Xi + α2Zit + εilt,

where αl is the market the listing is in, Xi is a vector of guest characteristics – whether the

guest is verified, if this was their 1st, 2nd, 3rd, or 4th+ trip, gender and age bin15 – and Zit is

a vector of characteristics – the date, number of guests, and (binned) number of nights – of

the trip taken by guest i in period t. For each listing we then calculate a listing effect

fl =1

nl

∑it

εilt =1

nl

∑it

(yilt+ − (α0 + α1Xi + α2Zit)) ,

where ml is the number of trips listing l has.

The guest return propensities will be imprecisely measured for listings with a small

number of trips hosted. For some of our analysis, we use only listings with at least 20

13This rate is by necessity larger than the overall average since any searches where we cannot identify theuser did not result in a booking.

14We could use a similar approach to try to measure guest quality, but since there are many fewer listingsthan guests, the number of guests with sufficiently many listings to get an accurate measure is fairly small.

15Gender and age are only observed for a subset of the guests. We use ‘unknown’ as a category for bothage and gender, but the results are similar if we use only guests for whom we observe these demographics.

15

Table 2: Summary search variables

(a) Searches (Jan 2016)

Statistic Mean St. Dev. Min Median Max N

New Traveler 0.76 0.43 0 1 1 28,114,802Days AdvanceSearched

77 79 −1 47 730 28,114,802

Num Nights 8 22 0 3 731 28,114,802Num Guests 3 3 0 2 16 28,114,802Entire Only 0.19 0.40 0 0 1 28,114,802Search ResultsNum Search Results 2,444 4,440 1 715 204,673 28,114,802Avg price of results 246 366 0 119 2,639 28,029,596Avg Jan 1 Rating 4.640 0.125 1.000 4.653 5.000 28,094,723Avg quality shown 0.170 0.192 −3.610 0.180 3.628 28,114,802OutcomesBooking (0/1) 0.035 0.183 0 0 1 28,114,802Trips After 1.753 3.309 0 1 339 17,628,029Spent After 637 1,677 0 80 642,443 17,628,029

(b) Booked Searches (Jan-June 2016)

Statistic Mean St. Dev. Min Median Max N

Quality booked 0.185 0.683 −3.610 0.286 3.628 4,454,159Rating booked 4.640 0.348 1.000 4.714 5.000 5,630,473Total price booked 365 569 0 206 60,300 7,129,432Trips After Check-in 1.939 4.030 0 1 338 6,771,328Spent After Check-in 571 2,319 0 97 1,525,804 6,771,328

Note: This table summarizes the search data. The top panel is all searches, the bottom panel is only those

searches which we are able to match to a booking. “New Traveler” includes searchers who are not linked

(by logging in or by the site recognizing their device) to an account or are linked to an account with zero

previous trips. “Days in advance” is the number of days between the search date and the searched for

check-in date. “Trips taken after” and “Dollars spent after” only include trips that are booked after the

search; these variables are only available for searchers who are linked to an account so we can calculate their

later behavior. The bottom panel summarizes characteristics that are only available for booked searches.

The first three lines give the quality, rating, and price of the trip linked to the search. The next lines give

the number of trips taken and number of dollars spent by those guests subsequent to the trip linked to the

search.

16

guest-trips on which to base the estimation. In addition, we adjust the listing effects using

a shrinkage estimator to account for the noise in the estimate (see Morris, 1983). To our

knowledge, prior work with shrinkage estimators has not accounted for the fact that, in many

contexts, one might expect the number of observations might be correlated with the effect

being estimated.16 The number of guests a listing has is likely not independent of its quality,

so it does not make sense to to shrink all effects towards the same mean; we modify the

usual application of shrinkage estimators accordingly. We group listings by the number of

trips they have and the time (quarter) that they were first created.17 If a group, g, has m

listings, and the average listing effect in the group is Eg[f ], then the weight for listing l with

nl trips is

δl = 1− m− 3

m

1

nl

Ei∈l[(εilt − fl)2]El∈g[(Eg[f ]− fl)2]

.

and the adjusted listing effect is

fl = δlfl + (1− δl)Eg[f ].,

Lastly, we winsorize the adjusted listing effects at 3 standard deviations from the mean to

get our final measure of GRP.

Figure 3a shows the distribution of the raw averages of the number of times a listing’s

guests return. It is limited to the 308k listings that had at least 20 trips prior to 2016. The

average across listings is 3.5 return trips.18 Figure 3b and 3c show the listing effects after

adding guest controls or guest and trip controls. Figure 3d shows the final the residualized,

shrunk, and winsorized listing effects, as described above. The standard deviation of the

raw effects is 1.87; after all controls it is 1.45; and for the final effects it is 0.55. Guest and

market controls do not decrease the variance much, the drop comes mostly from adjusting

for the noise from small samples. For the regression analysis below, we normalize the listing

effects to have a standard deviation of one (and a mean of zero).

Table 3 shows the correlations between our calculated listing effect and other metrics. The

adjusted listing effect is very highly correlated with the unwinsorized version and moderately

correlated with the raw averages. However, its correlation with other characteristics of the

listings is very low. The correlations with rating, price, number of guests, number of ratings,

16This seems potentially relevant for teacher quality – a common application of shrinkage estimators ineconomics.

17If two listings have the same number of guests, but one has been around twice as long, the newer oneis clearly the more attractive listing. Rather than trying to quantify this trade off, we group listings by ageinteracted with the number of guests.

18The average across guests is also 3.5 subsequent trips. Unlike the entire sample summarized in Table 1,these trips all happened prior to 2016, so the guests have had more time to accumulate additional trips.

17

0

20000

40000

60000

80000

0 5 10 15

Cou

nt

(a) Raw averages

0

20000

40000

60000

80000

−8 −4 0 4 8

Cou

nt

(b) Controlling for guest characteristics

0

20000

40000

60000

80000

−8 −4 0 4 8

Cou

nt

(c) Controlling for guest and trip characteristics

0

20000

40000

60000

80000

−8 −4 0 4 8

Cou

nt

(d) Residualized and shrunk effects

Figure 3: Guests’ return trips by listingNote: These histograms show the distribution across listings of the number of times their guests return, for

the 308k listings with at least 20 guests prior to 2016. Figure (a) is the raw averages. Figure (b) is the average

listing residual, controlling for guest characteristics; Figure (c) adds controls for the market and number of

guests on the trip. Figures (a), (b), and (c) censor 481, 837, and 747 observations, respectively. Figure (d)

shows the effects after shrinking towards a group mean to account for small samples, and winsorizing them

at 3 standard deviations from the mean.

and the share of guests who left ratings are all below 0.06. We do not know why the

correlation with ratings is so low. It may be that ratings are based on things guests think

are specific to a given listing, whereas whether a guest returns is not effected by things they

think are listing-specific, but precisely by the things they think generalize to all listings.19

Hosts on Airbnb can manage multiple listings. We compare the within host variance in

listing quality to the overall variance. Both overall and for listings with at least 20 guests,

the ratio of within host variance in quality to overall variance is about 85%, so we think that

quality is primarily a characteristic of the listing, rather than the host.

19It is also possible that for ratings guests have in mind some absolute standard rather than whether theythought the experience was good enough to try again. We thought the latter might be better captured bythe “value rating” that guests leave, but its correlation with quality is no higher.

18

Table 3: Correlation of quality measures across listings

Guest Return Propensity Unshrunk Raw Average

Guest Return Propensity 1 0.847 0.439Unshrunk 0.847 1 0.631Raw Average 0.439 0.631 1

Correlation of Listing Effect with:Rating Price Number of Guests Number of ratings #Ratings / #Guests

0.0478 0.0341 0.0216 0.0346 0.0509

Note: The top part of the table gives the correlation between the guest return propensity (our main measure

of quality), the unwinsorized version and the version not controlling for guest or trip characteristics, for the

308k listings with at least 20 guests prior to 2016. The bottom panel gives the correlation between the guest

return propensity (GRP) and the rating, the price, the number of guests, the number of reviews left, and

the fraction of guests who left a review. The price is the average price for October 2015 as of 09/01/15; the

correlation is based on the 241k listings for which both price and GRP is available. The other measures are

as of the end of 2015 and the correlations are based on the 292k listings for which those measures and GRP

are available.

3 ResultsThe guest return propensities (GRP) are calculated using the post-trip behavior of each

listing’s trips from 2011 through 2015. We look at the effect of GRP on subsequent guest

behavior using trips starting in 2016.

3.1 All trips

We focus on trips where the listing’s GRP is based on at least 20 trips. For each such

trip we calculate how many days until the guest either returns or ‘leaves the sample’ (the

data ends).20 Figure 4 shows hazard rates for a guest returning to Airbnb for listings in the

top and bottom quartiles of listing GRP; the hazard rates are the average probability that

a guest returns a given amount of time after staying at a listing, conditional on not having

returned between the stay at that listing and that point in time. At all time horizons the

probability of return is higher for guests that stayed at a listing in the top GRP quartile;

the difference falls by about half over the first 6 months.

Table 4 shows the regression results analogous to Figure 4. The first column is the raw

effect and the later columns control for ratings and guest and trip characteristics. These are

20For this part of the analysis, we exclude all trips where the subsequent trip was booked prior to the tripin question; though the guest could have cancelled that subsequent trip, we think that the effect of a listing’squality on whether the already-booked trip occurred would be negligible. For trips after the already bookedtrip, it would not be clear whether to use the initial listing or the one from the intervening trip.

19

0.003

0.006

0.009

1 2 3 4 5 6 7 8

Weeks

Ret

urn

Haz

ard

Bottom Quartile Top Quartile

(a) First 60 days

0.003

0.006

0.009

1 2 3 4 5 6 7 8 9 10 11 12 13

Months

Ret

urn

Haz

ard

Bottom Quartile Top Quartile

(b) First 400 days

Figure 4: Return hazard by quartile of listings’ guest return propensityNote: For listings in the top and bottom quartile of listings’ guest return propensity, these graphs show the

hazard rates of guests returning to Airbnb; the y-axis shows the (smoothed) probability that a guest returns

x weeks or months after staying with one of those listings, conditional on the guest not having returned

between the stay at that listing and time x. A listing’s GRP is measured by its guests prior to 2016, the

returns are shown for guests starting in 2016. Because we cannot control for market in the graphs, we use

a measure of GRP that does not control for market. The cyclicality in the top graph is because guests tend

to start trips on the same day of the week as their previous trip started on.20

the coefficients from Cox hazard regression;21 since guest return propensity is normalized

to have a standard deviation across listings of 1,22 the interpretation is that switching to a

listing with a 1 standard deviation higher GRP corresponds to a guest being exp(.098) ≈ 1.1

times as likely to return (at any given point in time, conditional on not having returned

previously). The remaining columns repeat the analysis separately for returns after a guest’s

first, second, third and fourth or more trips. The increase in return probability from a higher

GRP host is about 50% larger for first time guests than returning guests. This is consistent

with the idea that new guests know less about Airbnb so they update their priors more based

on the quality of the listing. The coefficient on rating bounces around, but even when it is

significant, it is qualitatively quite small.23

Table 4: Effect of listing guest return propensity on hazard of return

All Trips 1st 2nd 3rd 4+GRP Booked 0.237∗∗∗ 0.098∗∗∗ 0.123∗∗∗ 0.101∗∗∗ 0.096∗∗∗ 0.082∗∗∗

(0.001) (0.001) (0.003) (0.004) (0.004) (0.002)

Rating at booking −0.002 0.003 0.015∗∗∗ 0.006 −0.008∗∗

(0.002) (0.005) (0.006) (0.006) (0.003)

Controls 0 1 1 1 1 1Observations 9,003,856 5,773,355 1,668,324 1,074,219 740,012 2,290,800R2 0.004 0.134 0.114 0.066 0.047 0.067

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01Note: This tables show the coefficients from Cox Hazard regressions, analyzing the effect of listing GRP on

the probability that guest returns to Airbnb over time. Because we cannot include the thousands of market

fixed effects in the hazard regression, we use a measure of GRP that does not control for market. Column (2)

adds the listing’s rating at the time of booking and controls for when the initial trip happened, the guest’s

age bin, gender, experience, and verified-status as well as the trip’s number of nights, number of guests,

nightly price and whether it was booked through instant-book. Columns (3)-(7) run the same analysis as

Column (2) separately for different subsamples of guests – those for whom the trip was their 1st, 2nd, 3rd

or 4th or more trip on Airbnb.

The platform cares about how many times guest return and how much money they spend

21Since we cannot control for the thousands of market fixed effects in the Cox regression, for this analysis,we use a measure of guest return propensity that does not adjust for the listing’s market. If we use the fully-adjusted measure the effects are about half as large; however, we think our approach – controlling for marketin neither the guest return propensity calculations nor the regressions – is closer to controlling for market inboth analyses than this alternative of adjusting for market the guest return propensity calculations, but notthe regressions.

22The standard deviation across listings is 1, the standard deviation across trips in this sample is 0.4523Since the standard deviation of across listings of the rating is 0.58, the coefficient of 0.015 implies that

2nd time guests are 1.009 times as likely to return if they stay at a listing with a one standard deviationhigher rating.

21

●

●

●●

● ●●

●

●

●●

●●●

●

●●●

●

●

●●

●

● ●

●

●

●

●●

●

●●●

●

●●●

●

●

●

● ●

●●

●●

●●●

● ●

●

●

●

●

● ●

●●

●

●●

●

●●

●●

●

●

●●

●●

●

●

●●

●

●

●

● ●

●

●● ●

● ●●

●

●

●●●

●●

●

●

●

0.0

0.5

1.0

1.5

2.0

−1.0 −0.5 0.0 0.5 1.0

Listing Quality

Trip

s A

fter

(a) Trips after

●

●

●●

●●●

● ●

●

●

●● ●●

●●

● ●

●

● ●●● ●

●●

●

●● ●●●●

●● ●

●

●

●

●●

●

●●●● ●●●● ●

●

●●●

●●

●

●●

●●

●● ●

●●

●●

●● ●●●

●

●●●

● ●● ●

●

●●●

● ●●

●●

●●

●

●

●

●

●

●

0

100

200

300

400

500

−1.0 −0.5 0.0 0.5 1.0

Listing Quality

Spe

nt A

fter

(b) Spent after

Figure 5: Binscatter of trips taken and dollars spent by ventile of listing GRPNote: These graphs group listings into GRP percentiles. Each dot shows the average GRP of a group of

listings (x-axis) and the average number of trips taken (panel (a)) or dollars spent (panel (b)) by a guest

after staying at those listings. GRP is calculated based on trips prior to 2016. The y-axis measures are

based on trips starting in 2016. Because we cannot control for market in the graphs, we use a measure of

GRP that does not control for market.

on the platform, not just whether they return. For each trip we calculate how many trips

they booked (and took) after that trip and how many dollars were spent on those subsequent

trips. Figure 5 shows a binscatter of the trips taken and dollars spent after a trip by listing

GRP. These measures are also increasing in listing GRP, though the spending measure is

high variance.

Tables 5 and 6 give the corresponding regression results. On average, a 1 standard

deviation higher GRP host predicts guests taking 0.19 additional trips and spending $48

additional dollars on Airbnb. Here the effects tend to be larger for guests with more experi-

ence – because these guests travel more and spend more, even if though their beliefs about

Airbnb are less affected, small changes in their propensity to use it when traveling lead to

larger changes in the number of trips and dollars spent on the platform.24 The standard

deviation of rating across listings25 is 0.58 so the effect of a one standard deviation increase

in the rating is about a tenth of the effect of a standard deviation increase in GRP. The

effects for spending are more comparable.26

24Host quality could also affect the extent to which guests tell their friends about Airbnb and thereby“recruit” new potential guests. It seems likely that guests that are more likely to return are also more likelyto encourage friends to try Airbnb, so in general we think that if this is an important margin, we likelyunderstate the effects of host quality. However, the extent to which guests ‘recruit’ could also vary withguest experience – if experienced guests have already told all their friends, the effect of quality on new gueststhey recruit will be small. For a growing platform like Airbnb, a differential effect on new recruits could belarge enough to outweigh the differences in travel frequencies. With information on referrals, the platformcould adjust our metric to include effects on new guests.

25Like GRP, the standard deviation across trips is smaller than across all listings, it is 0.25.26Because the variance in spending is so high, we repeat the analysis with winsorized spending. The

22

Table 5: Effect of listing GRP on trips after

All Trips 1st 2nd 3rd 4+GRP Booked 0.011∗∗∗ 0.192∗∗∗ 0.070∗∗∗ 0.075∗∗∗ 0.084∗∗∗ 0.323∗∗∗

(0.002) (0.003) (0.003) (0.005) (0.006) (0.007)

Rating at booking 0.018∗∗∗ 0.026∗∗∗ 0.010 0.002 0.015(0.006) (0.006) (0.009) (0.012) (0.013)

All Controls 0 1 1 1 1 1Observations 11,499,195 7,710,665 2,016,255 1,384,152 989,703 3,320,555R2 0.00000 0.136 0.116 0.120 0.131 0.128

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Note: This table shows the effect of the GRP and rating of a listing on the number of trips a guest takes

subsequently. The first column looks only at the effect of GRP. The other columns add the listing rating

at the time the trip was booked and guest, trip and market controls. Columns (3)-(7) run the analysis

separately for guests for whom the trip was their 1st, 2nd, 3rd or 4th or more trip on Airbnb.

Table 6: Effect of listing GRP on spent after

All Trips 1st 2nd 3rd 4+GRP Booked 387.351∗∗∗ 47.980∗∗∗ 24.306∗∗∗ 12.976∗∗∗ 24.536∗∗∗ 73.378∗∗∗

(0.997) (1.591) (1.324) (3.144) (4.338) (3.082)

Rating at booking 26.475∗∗∗ 35.897∗∗∗ 26.580∗∗∗ 36.755∗∗∗ 20.921∗∗∗

(3.152) (2.580) (6.112) (8.469) (6.245)

All Controls 0 1 1 1 1 1Observations 11,499,195 7,710,665 2,016,255 1,384,152 989,703 3,320,555R2 −0.046 0.059 0.072 0.030 0.031 0.068

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Note: This tables show the effect of the GRP and rating of a listing on the amount of money a guest spend

on Airbnb trips subsequent to staying there. The first column looks only at the effect of GRP. The other

columns add the listing rating at the time the trip was booked as well as guest, trip and market controls.

Columns (3)-(6) run the analysis separately for guests for whom the trip was their 1st, 2nd, 3rd or 4th or

more trip on Airbnb.

23

Prices

The theory suggests that new information may also change guests’ beliefs about the

relationship between price and quality and therefore their optimal price point. We look at

whether high GRP has a different effect on guest behavior when it is a cheaper verse a more

expensive listing. First, we regress the nightly price of a trip on the number of guests, the

number of nights, and the market interacted with check-in date. We use the residual from

this regression as a measure of how expensive the listing was.27

Table 7 shows the effect of GRP by (residualized) price quartile, looking both at the

number of subsequent trips taken and the price of the next trip booked. Once we control

for guest and trip characteristics, there is only a slight difference in the effect of GRP by

price quartile, with GRP at lower priced listings having a slightly larger effect. We do

not see a significant difference in the effect of GRP by price quartile on the price of the

subsequent booking. This suggests that either guests do not update a lot their beliefs about

the correlation between price and quality or the different effects of correlation,28 shown in

Equation (1), cancel out.

If consumers do not properly account for the price level of a market or the added cost

of certain types of trips, then the relevant measure would be the actual price, not the price

residual. These results are shown in Appendix Table A2. The effect of GRP on subsequent

trips differs slightly more across quartiles (.20 for the lowest price quartile and 0.08 for the

highest price). The effect on prices is reversed: at a listing in the top price quartile, a higher

GRP leads guests to book a lower -priced subsequent trip, but the effect is tiny.29 A standard

deviation increase in GRP corresponds to a place that is $.78 cheaper per night (relative to

a standard deviation of $87).

results, shown in Appendix Table A1, are similar, though the effect for GRP is slightly smaller and that forlisting is somewhat larger. The largest effect of GRP is still on those with the most experience.

27We again exclude all trips where the subsequent trip was booked prior to the initial trip in question.See Footnote 20.

28Correlation increases the expected quality return to choosing a higher priced listing (pushing towardschoosing higher-priced listings) but also increases the expected quality at a given price, which pushes towardslower priced listings if there are decreasing returns to quality.

29As we discuss in Section 1.1, the choice to book a cheaper listing is consistent with the idea that there aredecreasing returns to quality and those listings are is some sense nicer than the guest wanted. The guest’sresponse is effectively, “if this is how good an expensive place is, than a moderately-priced place is probablyjust fine.” But the effect is so small, we do not put much stock in it.

24

Table 7: Effect of GRP by price quartile: price of next booking and subsequent trips

Trips Trips Price PricePrice Residual −0.0004∗∗∗ −0.0001∗∗ 0.234∗∗∗ 0.217∗∗∗

(0.00003) (0.00003) (0.001) (0.001)

Quality x Price Q1 −0.020∗∗∗ 0.147∗∗∗ 0.353∗∗∗ 0.094(0.005) (0.006) (0.125) (0.140)

Quality x Price Q2 −0.012∗∗∗ 0.125∗∗∗ 0.714∗∗∗ 0.173(0.005) (0.006) (0.121) (0.136)

Quality x Price Q3 0.038∗∗∗ 0.138∗∗∗ 0.796∗∗∗ 0.217(0.005) (0.006) (0.121) (0.137)

Quality x Price Q4 0.064∗∗∗ 0.116∗∗∗ 0.994∗∗∗ 0.470∗∗∗

(0.005) (0.006) (0.126) (0.140)

All Controls 0 1 0 1Observations 11,460,917 7,701,187 5,072,031 4,207,258R2 0.0005 0.247 0.071 0.084

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01Note: This tables show the effect of the GRP of a listing on the number of subsequent trips a guest takes

and the price residual of the next listing booked by a guest. The effect is shown separately by the quartile

of the price residual of the initial booking. The first and third columns looks only at the effect of GRP. The

second and fourth columns add guest, trip and market controls as well as the rating of the initial listing at

the time the trip was booked.

3.2 Searches

The evidence above suggests that these listing quality differences are persistent and mean-

ingful, but it is still possible that the difference is actually on the guest side – certain guests

are (unobservably) more likely to return and take more trips and also tend to chose certain

listings. To account for the endogeneity of guest choices, we use an instrumental variables

analysis based on guests’ search results. Frequent bookings, changes in availability, and

ongoing search experiments mean that guests searching for the same market and check-in

date on the same day are shown different search results. Their choice for where to stay is

affected by these results. At a macro scale, the results may be correlated with unobserved

guest characteristics – if some guests plan farther in advance and there may be higher-GRP

listings available then or certain types of guests may look to travel in certain markets that

have higher or lower average GRP. However, we believe that after controlling for the market

interacted with the check-in date searched for and the amount of time in advance of the

25

search, the time of day interacted with the timezone in which the search happened, and the

search filters, any residual variation in the GRP of the listings shown is as good as random.

Because the sample size is smaller, and the IV analysis is less affected by measurement

error, we use all listings for this analysis, not just those whose GRP is measured based on at

least 20 guests. Table 8 shows the reduced-form effects of the average GRP, rating, and price

on trips taken subsequently. The first two columns of Table 8 show the reduced-form effect

of the average GRP and rating of the listings shown on the number of trips the searcher

takes after searching. The first column is for all searches we can identify; the second is for

those who book a trip related to the search – the sample relevant for the IV analysis. The

first column uses searches from January, 2016. When we limit to booked searches the sample

size shrinks dramatically, so we use searches January – June 2016.30 Columns (3) and (4)

use the same sample of booked searches as Column (2) but look at the effect on the number

of trips the searcher takes after check-in. For the IV, where we’re interested in the effect of

the listing characteristics, it makes more sense to use this measure, which does not include

trips taken prior to the trip at that listing.31 Column (3) uses the number of past trips the

guest had taken at time of search as a control, Column (4) uses the number of past trips

the guest had taken at the time of check-in – again the variable more relevant for the IV

analysis. All columns include the full set of market, check-in, and search timing controls.

Recall that GRP is normalized to have a standard deviation of 1 across listings. However,

the standard deviation in the average GRP of the search results is only 0.19. So Column (4)

of Table 8 implies that raising the GRP of all the listings by one standard deviation of listing

GRP results in 0.14 more trips, but raising the average GRP by one standard deviation of

average GRP shown only results in 0.03 additional trips. The effects of rating are even

smaller. The standard deviation across listings in the rating as of January 1 is 0.58 and the

standard deviation in average rating shown is 0.13. So raising the rating of all the listings

by one standard deviation of listing rating would result in 0.05 more trips, but raising the

average rating shown by one standard deviation of average rating shown only results in 0.008

additional trips.

One reason why the search results have a fairly small effect on subsequent behavior is

that their effect on the type of listing booked is modest. The first two columns of Table

9 show the effect of the average GRP and rating shown on the GRP and rating booked.

30Since our quality measure is as of the end of 2015, the fraction of search results with a quality measureavailable decreases over time, as the fraction of listings created since the beginning of January, 2016 increases.To the extent that quality changes over time, the quality measure also becomes less accurate. For thesereasons, we use only the first half of 2016.

31Trips would be include in the first measure and not the second if they were booked after the search andtook place before the associate booking.

26

Table 8: Effect of average GRP and rating of search results on number of subsequent trips

Trips After Search Trips After check-in

JanuarySearches Booked Searches

Avg quality shown 0.067∗∗∗ 0.118∗∗∗ 0.125∗∗∗ 0.139∗∗∗

(0.005) (0.015) (0.014) (0.014)

Avg Jan 1 Rating −0.046∗∗∗ 0.046∗∗ 0.032 0.063∗∗∗

(0.008) (0.022) (0.021) (0.021)

Avg price of results −0.00000 −0.0001∗∗∗ −0.00004∗∗∗ −0.00004∗∗

(0.00000) (0.00002) (0.00001) (0.00001)

1 Past Trips 0.833∗∗∗ 0.727∗∗∗ 0.591∗∗∗

(0.002) (0.006) (0.005)

2 Past Trips 1.519∗∗∗ 1.271∗∗∗ 1.036∗∗∗

(0.003) (0.007) (0.007)

3+ Past Trips 3.524∗∗∗ 3.612∗∗∗ 3.038∗∗∗

(0.002) (0.005) (0.005)

1 Past Trips 0.568∗∗∗

(0.005)

2 Past Trips 0.993∗∗∗

(0.006)

3+ Past Trips 2.874∗∗∗

(0.005)

Observations 16,779,240 6,964,635 6,616,009 6,616,009R2 0.256 0.482 0.486 0.485Adjusted R2 0.172 0.162 0.157 0.154∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Note: The first column shows the effect of the average GRP and rating of search results on the number of

trips taken afterwards for all guests we can identify for searches in January 2016. The remaining columns

are limited to searches that resulted in a booking; to maintain a large enough sample we use searches in

January - June 2016. Column (2) repeats the analysis of Column (1) of trips taken after the search for the

IV-subsample. Columns (3) and (4) show the effect of the average GRP and rating of search results on the

number of trips taken after the check-in date. For the first three columns, “Past Trips” refers to the number

of trips the user took prior to the search; for the last column it is the number of trips prior to the trip’s

check-in date; for both zero is the reference group. All regressions control for market x check-in-date x how

far in advance the search happened, the search date, the hour of the search interacted with the timezone the

searcher was searching from and any search filters the searcher used.

27

The estimated effect of average GRP shown on average GRP booked is 0.44; for ratings the

analogous effect is 0.33. The cross effects are negligible. Our measures of the characteristics

of the listings shown use only the average of the results on the first page and only the first

search the user makes for a given market and check-in range (to avoid guests’ desire to search

more influencing the search results); other results they see may differ, so the effect of the

listings shown on the listing booked is less than 1.

The last column of Table 9 puts the preceding results together in an IV regression. We

instrument for the characteristics of the booked listing with the characteristics of the search

results and look at the effect on trips taken afterwards. We find that a one standard deviation

increase in the GRP of the listing booked leads to 0.34 additional trips by that guest. On

average, in this sample, a guest takes 2 subsequent trips, so that is an increase of 17%.32 The

coefficient on the rating is smaller and recall that there is also less variance across listings

in their rating so a one standard deviation increase in the rating of a listing leads to 0.13

additional trips by the guest.33

The identifying assumption for our IV regressions is that given the controls – (1) the

market searched for, interacted with the check-in date searched for, interacted with bins

for how far in advance the search was34 (2) the date the search happened (3) the timezone

the searcher was in interacted with the hour of day the search happened (4) the search

filters the searcher used – the remaining variation in the GRP and rating of listings shown is

uncorrelated with guest characteristics. The remaining variation results from listings being

added, listings being booked, and experiments in Airbnb’s search algorithm.35

In addition to affecting what place a searcher books if the search results in a booking,

the search results can also affect whether the search results in a booking. Table 10 looks

at the effect of search results on this latter, extensive margin effect. The effects are fairly

small: a 1 standard deviation increase in the average GRP shown increases the booking rate

by .002 ∗ .19 ≈ .00038 percentage points. Since the baseline booking rate for known guests

is about 5.6%, this an increase of 0.68%. The fact that these effects are non-zero means

32The 0.34 coefficient is not directly comparable to the standard deviation of 0.55 trips that GRP hadbefore we normalized it to 1, because it is based on an earlier sample. The mean number of return trips fortrips prior to 2016 was 3.5.

33If instead of controlling for the average price shown, we instrument for the price booked with the priceshown, the coefficients on GRP and rating are largely unaffected and coefficient on price is still tiny.

34We do separate bins for each of 0 through 30, then bin by week for searches less than a year in advance.For searches more than a year in advance we have two bins: less than 400 and greater than 400. Since wehave thousands of markets and thousands of check-in dates, this binning makes the interaction tractable.The exact search date is also controlled for, but not interacted with the market and check-in date.

35We do not directly use this experimental variation because the first stage F-statistic is about 1. Sincethe search algorithm is not considering this measure of GRP, and it is not strongly correlated with prices orratings, the experiments do not have large effects on the average GRP shown. We are working with Airbnbto implement an experiment that directly varies the GRP of listings shown.

28

Table 9: First stage and IV results:Effect of search results on booking characteristics, and instrumented effect of booking

characteristics on subsequent trips

First stage

GRPBooked

RatingBooked

Trips After(IV)

Quality booked 0.340∗∗∗

(0.042)

Rating booked 0.218∗∗

(0.089)

Avg quality shown 0.439∗∗∗ −0.002(0.003) (0.001)

Avg Jan 1 Rating 0.003 0.325∗∗∗

(0.005) (0.002)

Avg price of results −0.00003∗∗∗ −0.00001∗∗∗ 0.00000(0.00000) (0.00000) (0.00002)

1 Past Trips 0.002 0.009∗∗∗ 0.571∗∗∗

(0.001) (0.001) (0.007)

2 Past Trips 0.001 0.014∗∗∗ 0.988∗∗∗

(0.002) (0.001) (0.008)

3+ Past Trips 0.005∗∗∗ 0.025∗∗∗ 2.851∗∗∗

(0.001) (0.0005) (0.007)

F-Stat . . 33438.4Observations 4,357,947 5,610,820 4,133,270R2 0.507 0.451 0.548Adjusted R2 0.089 0.055 0.151∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Note: The first three columns show the effect of search results on the characteristics of the booking, the first

stage for the IV regression. The second column is limited to trips at listings for whom the GRP measure is

based on at least 20 trips. The last column shows the IV results where we use the average GRP and rating

of search results to instrument for the GRP and rating booked. For the first three columns, “Past Trips”

refers to the number of trips the user took prior to the search; for the last column it is the number of trips

prior to the trip’s check-in date; for both zero is the reference group. All regressions control for market x

check-in-date x how far in advance the search happened, the search date, the hour of the search interacted

with the timezone the searcher was searching from and any search filters the searcher used.

29

Table 10: Effect of search results on whether a booking occurs

All AllAvg quality shown 0.001∗∗∗ 0.002∗∗∗

(0.0002) (0.0004)

Avg Jan 1 Rating 0.003∗∗∗ 0.003∗∗∗

(0.0004) (0.001)

Avg price of results −0.00001∗∗∗ −0.00002∗∗∗

(0.00000) (0.00000)

1 Past Trips 0.007∗∗∗ 0.007∗∗∗

(0.0001) (0.0002)

2 Past Trips 0.017∗∗∗ 0.016∗∗∗

(0.0002) (0.0002)

3+ Past Trips 0.036∗∗∗ 0.035∗∗∗

(0.0001) (0.0002)

Unknown Guest −0.048∗∗∗

(0.0001)

Observations 26,895,821 16,779,240R2 0.096 0.112Adjusted R2 0.032 0.012∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

Note: This table shows the effect of the average GRP and rating shown on whether a search results in

a booking. Searches are matched to bookings by that user that happen within 6 days of the search and

have the same destination market and check-in date +/− 1 day. The first two columns include searches by

unidentified searchers (none of which result in a booking). In the second column the excluded group for the

number of past trips is unknown searchers (whose number of past trips is unknown). The last column is

limited to identified searchers.

that when we look at the affect of booking characteristics on subsequent guest behavior, we

have a slightly selected sample. However, if we assume that the guests who were less likely

to book (and only did so because there were listings with a higher rating or GRP) are also

less likely to take additional trips, then our results are a lower bound on the effect on an

unselected sample.

Guest Experience

Guests with different amounts of past experience with Airbnb will both have more infor-

mation about the platform and are likely to be people who travel more frequently. The first

column of Table 11 repeats the IV analysis from the third column of Table 8; the subsequent

columns show the same regression separately for guests for whom this is their first, second,

third, or fourth (or more) trip. The effect of GRP is larger for guests with more experience

30

on Airbnb: the higher frequency of travel outweighs the smaller learning effect.36 Interest-

ingly, though we lose power, the point estimates suggest that the effect of rating might be

largest for first time guests.

Table 11: Effect of GRP and rating booked on subsequent trips, by guest experience (IV)

By Trip NumberAll 1st 2nd 3rd 4+

Quality booked 0.340∗∗∗ 0.093∗∗ 0.195∗ 0.271 0.946∗∗∗

(0.042) (0.038) (0.105) (0.188) (0.198)

Rating booked 0.218∗∗ 0.264∗∗∗ 0.095 0.611 −0.627(0.089) (0.076) (0.213) (0.396) (0.411)

Avg price of results 0.00000 0.00003 0.00001 0.0001∗ 0.00000(0.00002) (0.00002) (0.00005) (0.0001) (0.0001)

F-Stat 33438.4 14208.5 5406.2 3383.4 7474.9Observations 4,133,270 1,800,689 764,901 446,175 1,112,354R2 0.548 0.621 0.770 0.838 0.700Adjusted R2 0.151 0.055 0.070 0.062 −0.004

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01Note: This table repeats the IV analysis of the effect of listing GRP on subsequent guest trips for subsamples

of guests based on their number of prior trips with Airbnb. The first column repeats the results from the

last column of Table 9, for reference. Columns (2)-(5) do the same analysis for guests for whom it is their

first, second, third or fourth (or more) trip. All regressions control for market x check-in-date x how far in

advance the search happened, the search date, the hour of day of the search interacted with the timezone

the searcher was searching from and any search filters the searcher used.

What about Dollars?

If the platform takes a percentage cut instead of a flat fee, it will care about the dollars

spent by guests, not just the number of times they return. Table 12 looks at the effect of the

GRP and rating booked on the amount (in dollars) that the guest spends on subsequent trips

on Airbnb, again instrumenting for the GRP and rating booked with the average GRP and

average rating of the search results. Here rating plays a larger role, but the effect of GRP is

still significant – statistically and economically. A one standard deviation increase in GRP

leads to $111 more spent subsequently. The standard deviation across listings of the rating

36In this sample a first-time guest returns an average of 0.9 times and a fourth-time guest returns anaverage of 4 times, so the point estimates suggest that, contrary to our expectations, the percentage increaseis also larger for more experienced guests, but the estimates are not precise enough to rule out the opposite.

31

is 0.58 so a standard deviation increase in rating leads to $250 more spent subsequently,

slightly larger than the effect of GRP.37

Table 12: Effect of GRP and rating booked on subsequent spending, by guest experience(IV)

By Trip NumberAll 1st 2nd 3rd 4+

Quality booked 106.310∗∗∗ 35.693∗∗ 40.147 119.733 382.154∗∗∗

(22.090) (17.127) (49.334) (86.658) (102.008)

Rating booked 425.567∗∗∗ 160.328∗∗∗ 203.124∗∗ 430.627∗∗ 707.292∗∗∗

(47.360) (33.930) (100.633) (182.759) (211.877)

Avg price of results 0.327∗∗∗ 0.157∗∗∗ 0.214∗∗∗ 0.415∗∗∗ 0.661∗∗∗

(0.011) (0.008) (0.022) (0.041) (0.043)

F-Stat 33438.4 14208.5 5406.2 3383.4 7474.9Observations 4,133,270 1,800,689 764,901 446,175 1,112,354R2 0.465 0.570 0.718 0.810 0.716Adjusted R2 −0.006 −0.073 −0.139 −0.105 0.050

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01Note: This table uses the same IV approach as Table 11 to show the effect of listing characteristics on a

guest’s subsequent spending (in dollars) on the platform. The first column shows the effects for all guests

(controlling for guest experience). Columns (2)-(5) do the same analysis for guests for whom it is their first,

second, third or fourth (or more) trip. All regressions control for market x check-in-date x how far in advance

the search happened, the search date, the hour of the search interacted with the timezone the searcher was

searching from and any search filters the searcher used.

Since higher rated listings are more expensive (and higher GRP ones are not), it is

perhaps not surprising that rating has a large effect on dollars spent even though GRP has

a larger effect on trips. If booking more highly rated places also causes guests to return

to more highly rated, and therefore more expensive places, the effect on spending would be

disproportionately larger than the effect on the number of trips. 38

37A version of Table 12 using winsorized spending is Appendix Table A3; the results are similar.38In Appendix Table A4, we look at this directly by using the price and rating of the place booked

subsequent to the trip of the booked search as the left-hand side variable. The sample size is smaller becausethis only includes booked searches where the guest made an additional booking after the trip associated withthe search. As expected, the rating booked increases both the rating and the price of the subsequent listingbooked. The effect of GRP is negligible. The price initial listing has a small effect on the price of the nextbooking.

32

4 Implications for the PlatformHaving seen that there is substantial variation across listings in the externality that they

exert on the rest of the platform, the next question is what the platform can do with that

information. We consider four ways the platform can raise listing quality or redirect guests to

higher-quality listings.39 Often when there are externalities in a market, the efficient solution

is a Pigouvian tax to make agents internalize the externality. However, platforms also have

the option to incorporate quality into the ranking of search results, to inform users about

listing quality, to remove low quality hosts from the platform, or to try to help low-quality

hosts improve.

Taxing the externality

The platform could pay hosts when their guests return or fine them if the guests do not

return within a certain time.40 This would help high-quality hosts over low-quality hosts,

and give all hosts an incentive to raise their quality. Since guests’ returns are stochastic, the

tax would also create additional variance in a host’s revenue.

How much host quality would improve depends on how convex the costs to improving

quality are, where costs include both the cost of figuring out what things make guests return

and the cost of doing those things. Unless the platform could also give hosts guidance as to

how to improve their guest return propensity, the changes might be fairly small. However,

even if the externality for a given transaction is unchanged, Pigouvian taxes can improve

efficiency by affecting prices (and quantities). A revenue-neutral tax/subsidy for quality

would cause low quality hosts to increase prices and high quality host to decrease prices,

shifting demand to hosts with better externalities.41

To get a sense of the magnitude of the effects of a Pigouvian tax on quantities and surplus,

imagine an incremental implementation of a tax/subsidy in proportion to quality. If quality

is measured in trips, and each trip has a value to the platform as a whole (including sellers)

of V , then the revenue-neutral tax would be τj = τ(GRP − GRPj)V . If quality does not

change, the change in platform surplus is the change in each host’s number of guests, ∆dj,

39We are implicitly assuming that a listing’s quality is not affected by the number of guests it gets. Ourestimates cannot speak to whether average and marginal quality are the same – if a listing gets more guestsit could gain experience and improve or be too busy and lose quality. Nevertheless, we think the platformwill likely want guests to go to higher-quality listings.

40It would probably be more practical to give the host part of the payment at the time of the trip and therest only if the guest returns.

41If there are fix costs to being on the platform, a tax on low quality could also cause low-quality hosts toleave the platform. This is related to the effects of screening discussed below.

33

times the number of trips a guest who stays with that host takes, times the value of a trip:42

∆S =∑j

∆dj(1 +GRPj)V.

The changes in quantities will depend on how much hosts pass through their cost changes

to prices and on how responsive demand is to price. If we take a super-simplified case where

the equilibrium prior to the tax is symmetric, then marginal costs, cj, the own- and cross-

derivatives of demand(∂dj∂pj,∂dj∂pk

), and own- and cross- pass-through rates

(∂pj∂cj,∂pj∂ck

), will all

be the the same across listings. In this case, the effect on welfare is

∂S

∂τ= −J · ∂dj

∂pj(1 +DRjk)

(∂pj∂cj− ∂pj∂ck

)var(GRP ) · V 2, (2)

where J is the number of listings and DRjk = −∂dk∂pj

/∂dj∂pj

is the diversion ratio (the fraction

of the consumers that j loses when it raises its price that become consumers of k).43

The tax is revenue neutral, so average cost change across all listings is zero. Therefore,

as the number of listings in the relevant market gets larger, the average change in the

competitors’ costs is close to zero. Similarly, the changes in competitors’ prices will cancel

out. Because of this, the effect of demand diversion (DRjk) and cross pass-through (∂Pi

∂cj)

will be negligible. Moreover, if sellers are pricing optimally then the margin equals the semi-

elasticity of demand Vpj

= pj − cj =dj∂dj∂pj

. In this case, the fraction increase in surplus is

approximately44

∂S

∂τ

1

S≈ ∂pj∂cj

var(GRP ). (3)

We see from Equation (2) when the benefits of a Pigouvian tax will tend to be larger.

When demand is more responsive to price (for a given number of listings), more consumers

will shift to higher quality listings. Therefore, the tax is more beneficial when demand is

more responsive. Similarly, when more consumers switch to other listings rather than leaving

the platform when a price increases, (i.e. the diversion ratio is higher), the benefit from the

tax is larger. In addition, the more listings pass the cost change through to prices, the more

42If guests who do not book a given trip still return in the future, then this term should in include∆d0(1 +GRP0), where ∆d0 = −

∑j ∆dj

43As τ moves away from zero, the equilibrium becomes less symmetric, so this formula holds for all τ onlyif demand is linear and pass-through is constant over the relevant range of costs.

44We simplify

∂S

∂τ≈ −J · ∂dj

∂pjV · ∂pj

∂cjvar(GRP ) · V = −J · dj

∂pj∂cj

var(GRP ) · V,

and use the fact that total surplus is the number of trips J · dj times the surplus per trip V .

34

the tax will affect prices and shift demand; so, the benefit from the tax is higher when pass-

through is higher.45 Lastly, and unsurprisingly, the welfare gains from the tax are larger

when there is higher variance in quality and higher value of each trip.

For a back-of-the-envelope calculation, we can use Equation (3) and our estimate of 0.34

for the effect of a standard deviation increase in quality on guests’ subsequent trips. If a

listings’ pass-through rate were one-half,46 and implementation and brand-image concerns

were not an issue, a Pigouvian tax on quality could increase the surplus per trip by about

6%.

Search results

While the platform may consider Pigouvian taxes infeasible or undiplomatic, the platform

can directly shift quantity from low to high quality hosts, without taxes or price changes,

via its search algorithm.47 Like taxes, including quality in the search rankings would give

hosts an incentive to improve quality, if they are aware that it affects the search ranking.

Promoting high-quality listings would also raise the surplus for given levels of quality. The

more search rankings affect which listings guests book, the more incorporating quality into

the search algorithm will raise efficiency and incentivize hosts to improve.

To think about the platform’s trade offs in which listings to promote to users (e.g. put on

the first page of results) we need to think about another dimension of listing quality. Listings

vary not only in the propensity of their guests to return to the platform, but also in their

propensity to be booked when shown to guests – their purchase propensity or ‘attractiveness’.

In reality, there is horizontal heterogeneity – a listing’s booking rate will depend on which

other listings it is shown with. Listings may also vary in what fraction of their bookings are

‘new’ bookings as opposed to bookings ‘taken’ from other listings. However, since our focus

is on the effect of incorporating guest return propensity into the rankings, we assume that,

conditional on what the platform knows about a searcher, attractiveness is uni-dimensional.

If the platform is just trying to maximize current bookings, it will show the most at-

tractive listings (those with the highest booking-propensity) that meet the search criteria.

If the platform is maximizing the number of bookings in the long term, it will trade-off the

listing’s attractiveness, A, and its guest return propensity, GRP . Abstracting from guest

45Since a listing’s cost moves in the opposite direction as the average of other listings’ costs, if cross pricepass-through is positive, that attenuates the listing’s price change.

46If we do not think sellers are necessarily pricing optimally, we can choose the elasticity of demand andmargin separately. If ε = −4, and V

P = .2, then the fractional gain in surplus gain is 4 · .2 · .5 · .12· = .05·.The formula is linear in the elasticity, the margin, and the pass-through rate, so it is easy to see how theresult would change for different values of these parameters.

47Changes in the search algorithm can be thought of as the ‘nudge’ alternative to the ‘paternalistic’ banningof hosts.

35

heterogeneity, the platform will pick the listings with the highest total expected bookings

b(Al, GRPl) = Al(1 +GRPl −GRP0),

where GRPl is the expected number of trips that a guest takes after staying with that host,

and GRP0 is the number of trips that a guest who does not book this trip will end up taking

in the future.48

The gains from maximizing b(Al, GRPl) instead of just Al are (1) decreasing in the

correlation between GRP and A, (2) increasing in the variance of GRP , and (3) decreasing

in the variance of A.49 If we use ratings as a proxy for attractiveness, then in the case of

Airbnb the correlation between A and GRP is low. The variance in GRP is about 0.11

trips. Our data do not speak directly to the variance of A (since ratings are not in the units

of booking probability). However, if we think about the gain from re-ordering only the 20

most attractive listings by the true ranking b(Al, GRPl), we can say that the variance of

A among those 20 is decreasing in the total number of listings in the market (as long as

the distribution of A is log-concave).50 So the value to paying attention to GRP as well as

attractiveness is increasing in the number of listings in a market.

In a large market, if the top 20 listings have about the same attractiveness, then the value

of re-ordering those listings would depend just on the variance of GRP and the amount that

booking probability changes with ranking. If GRP is normally distributed, the top 10 values

will, on average, be 0.77 standard deviations above the mean (and the bottom half will be 0.77

s.d. below). Ursu (2016) find that Expedia results ranked 1-10 are about 4 times as likely to

be booked as those ranked 11-20. If on average listings in the top half are four times as likely

to be booked, the optimal order would result in an additional .77 · (4− 1)/(4 + 1) · .34 ≈ 0.16

trips per booking.51 If we use 2 or 5 times as likely, instead of 4, we get 0.09 or 0.17 trips,

respectively.

48This manipulation of search results is in the interest of guests, helping them book the unobservablyhigher quality listings. On the host side it will help some hosts over others, but it improves efficiency in thesense that it puts hosts’ individual incentives more inline with the social surplus.

49If the platform were just choosing the maximum A, then in expectation it would get

E[b(Al, GRPl)] = E[A](1 + µGRP ) + ρ · var[A]

where µGRP is the mean of GRPl, unconditional on Al and ρ is the correlation between Al and GRPl.50We can think of the top 20 as being drawn from a truncated distribution with a truncation point at the

21st best. As long as the distribution of A is log-concave (as is the case for most common distributions), thevariance of the truncated distribution is decreasing in the truncation point. Increasing the total number oflistings leads to a first-order stochastically dominate distribution of the 21st best draw, so the variance ofthe distribution of the top 20 is necessarily decreasing.

51If only the top 10 listings had equivalent A, we care about the top 5 verse the next 5. Ursu (2016) findthe top 5 are about 2.5 times as likely to be booked as the next five; the expected GRP of the top 5 is 0.74standard deviations above the mean, so the gain would be .74 · (2.5− 1)/(2.5 + 1) · 0.34 ≈ 0.11 trips.

36

The average searcher in our sample took 2 trips subsequent to their booked search. The

gain to reordering their search results would depend on the variance in booking probability

and its correlation with GRP . For a large market, our ballpark calculations suggest incorpo-

rating the guest return propensity into search ranking could generate an increase in return

trips on the order of 4.5-8.5%.

Informing guests

In addition to or instead of altering the search results based on GRP, the platform could

report GRP to users, the way it does with ratings. However, it might be hard to explain to

guests what GRP is measuring. The platform could show guests the GRP as an undefined

‘platform quality’ measure, but that might frustrate hosts who feel it is out of their control.

The risk with presenting the information – in whatever form – is that doing so will change

guests’ expectations. To the extent that what is important for returns is how good a listing

is relative to the guest’s expectations, telling guests a listing’s quality could have substantial

unintended consequences.

Screening and Improving

Airbnb also has the option to remove low-quality hosts. The value of this will depend on

the thickness of the platform-market and also on the extent and type of competition Airbnb

has in the broader market. In our model of consumer learning, we assumed that the value

of the outside option was known. If instead, the outside option is (partially) other peer-to-

peer accommodation platforms, guests may update their priors not just about the quality

of Airbnb, but about the quality of all such platforms. This changes Airbnb’s incentives.

It means they want hosts on all platforms to be high quality, so that people return to the

peer-to-peer accommodation market as a whole. If Airbnb kicks off a bad host, that host

may register on another platform.52 That is still not good for Airbnb. So if a platform

thinks that users are learning about the industry as a whole, they have an incentive to try

to improve listing quality, rather than just remove low-quality listings.

Conversely, if the market as a whole is more mature or well-known, Airbnb may think

that users are only learning about the quality of sellers on its platform. In that case, it has

much more of an incentive to remove bad hosts. Not only is it potentially easier to screen

hosts than improve them, but if a bad host went to another platform, that could be good

for Airbnb. Diverting low-quality hosts to another platform increases the probability that

guests conclude that the other platform is low quality, and switch to Airbnb.

52If hosts are multi-homing (selling through multiple platforms), then a host kicked of Airbnb could increaseits number of transactions on another platform.

37

5 ConclusionWe propose measuring sellers’ quality on a platform by the externality they impose on

other sellers on the platform – how many times consumers who purchase from a given seller

return to the platform. Guest return propensity is a ‘revealed preference’ measure of quality,

based on what consumers do after after purchasing from a seller, rather than what they say

about the seller. For Airbnb listings with at least 20 guests prior to 2016, the raw standard

deviation across listings in the number of times their guests return is about 1.9; the average

is 3.4. After accounting for observable guest and trip characteristics and adjusting for small

samples, the estimated standard deviation of GRP is 0.55 trips. Our IV estimates, using

variation in users’ search results, suggest that a standard deviation increase in a listing’s

GRP causes a an increase in a guest’s subsequent trips of 0.34 (17%).

We also highlight an important aspect of consumer heterogeneity. Even though we believe

inexperienced buyers learn more about the platform, in many contexts, experienced buyers

will be frequent buyers. The lower probability of an experienced guest leaving the platform

due to one bad experience may be outweighed by the much higher cost to the platform if

that guest leaves. This is analogous to the airline reward programs. Though experienced

frequent flyers are probably less likely to switch airlines because of one bad flight, if they do,

the revenue loss is much greater than if an infrequent consumer switches.

The gains to other platforms from incorporating this quality measure will vary. The

value of screening bad sellers will depend on the broader market that the platform operates

in. When quality is very observable ex-ante, we would expect our measure of quality to

not vary much and potentially be correlated with the measures the platform already uses

for evaluation – such as propensity to purchase. The externalities are smaller in markets

where consumers have strong priors or believe there to be a lot of variance across sellers,

because buyers will change their beliefs less based on a given experience. By contrast, in

thick markets, with lots of good options, platforms may be able to nudge users towards sellers

who are higher-quality by our metric, and not particularly lower quality by other metrics.

The importance of potential return business will also matter;53 the more important return

users are to a platform’s business model, the more weight should be given to our measure of

quality relative to purchase propensity.

53A platform for wedding photographers, for example, probably does not have a lot of return business;while there still may be an externality in terms of whether people recommend it to their friends, it cannot becaptured with our measure. Frequent consumers who are not myopic will value learning about a platform, sotheir threshold for not returning may be lower, but the cost to the platform of a seller pushing them acrossthat threshold is higher.

38

Appendix: Supplementary Tables

Table A1: Effect of listing GRP and rating on subsequent spending, by guest experience

All All 1st 2nd 3rd 4+Quality Booked 346.285∗∗∗ 34.832∗∗∗ 21.030∗∗∗ 16.154∗∗∗ 19.720∗∗∗ 49.229∗∗∗

(0.515) (0.710) (0.998) (1.368) (1.754) (1.300)Rating at booking 43.736∗∗∗ 36.018∗∗∗ 35.982∗∗∗ 40.270∗∗∗ 54.824∗∗∗

(1.406) (1.945) (2.660) (3.425) (2.634)

All Controls 0 1 1 1 1 1Observations 11,499,195 7,710,665 2,016,255 1,384,152 989,703 3,320,555R2 −0.152 0.164 0.102 0.115 0.130 0.177

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01Note: This tables uses winsorized spending and replicates the analysis from Table 6 of the effect of the GRP

and rating of a listing on the dollars a guest spent on Airbnb after staying there. The first column looks only

at the effect of GRP. The other columns add the listing rating and full set of controls. Columns (3)-(7) run

the analysis separately for guests for whom the trip was their 1st, 2nd, 3rd or 4th or more trip on Airbnb.

Table A2: Effect of GRP by price residual quartile on price of next booking and trips taken

Trips Trips Price PriceNightly Price −0.001∗∗∗ −0.0004∗∗∗ 0.249∗∗∗ 0.281∗∗∗

(0.00002) (0.00003) (0.001) (0.001)Quality x Price Q1 0.092∗∗∗ 0.197∗∗∗ 1.031∗∗∗ 0.146

(0.004) (0.006) (0.146) (0.166)Quality x Price Q2 0.030∗∗∗ 0.120∗∗∗ −0.252 −0.427∗∗

(0.005) (0.006) (0.166) (0.186)Quality x Price Q3 0.020∗∗∗ 0.107∗∗∗ −0.977∗∗∗ −0.470∗∗

(0.005) (0.006) (0.172) (0.192)Quality x Price Q4 0.013∗∗∗ 0.081∗∗∗ −2.321∗∗∗ −0.775∗∗∗

(0.005) (0.006) (0.179) (0.200)

All Controls 0 1 0 1Observations 11,498,914 7,710,559 5,107,111 4,227,932R2 0.007 0.247 0.138 0.150

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01Note: This tables show the effect of the GRP of a listing on the price residual of the next listing booked

by a guest and the number of subsequent trips that guest takes. (Only including initial trips where the

subsequent trip was booked after the check-in date of the initial trip.) The effect is shown separately by the

quartile of the price residual of the initial booking. The first and third columns looks only at the effect of

GRP. The second and fourth columns add guest, trip and market controls as well as the rating of the initial

listing at the time the trip was booked.

39

Table A3: IV:Effect of GRP and rating booked on winsorized subsequent spending

By Trip NumberAll 1st 2nd 3rd 4+

Quality booked 99.376∗∗∗ 19.945 64.081 79.651 285.970∗∗∗

(12.549) (14.791) (39.797) (67.966) (51.412)Rating booked 396.961∗∗∗ 168.318∗∗∗ 245.938∗∗∗ 580.355∗∗∗ 786.608∗∗∗

(27.224) (30.723) (82.837) (150.923) (111.134)Avg price of results 0.462∗∗∗ 0.244∗∗∗ 0.450∗∗∗ 0.745∗∗∗ 0.899∗∗∗

(0.012) (0.014) (0.035) (0.064) (0.044)

F-Stat 32741.7 13884.8 5237.7 3500 7347Observations 3,869,881 1,626,456 729,047 429,669 1,084,709R2 0.520 0.604 0.744 0.821 0.693Adjusted R2 0.075 −0.040 −0.068 −0.072 −0.043

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01Note: This table replicates the analysis of the effect of listing characteristics on a guest’s subsequent spending

(in dollars) on the platform from Table 12, with a winsorized measure of subsequent spending. Each column

is a subgroup of guests based on whether it was their first, second, third or fourth (or more) trip. All

regressions control for market x check-in-date x how far in advance the search happened, the search date,

the hour of the search interacted with the timezone the searcher was searching from and any search filters

the searcher used.

Table A4: Effect of listing characteristics on rating and price of next trip booked (IV)

Rating of Next Booking Price of Next BookingQuality booked 0.005 0.639

(0.006) (1.226)

Rating booked 0.210∗∗∗ 37.915∗∗∗

(0.014) (2.650)

Avg price of results 0.00000 0.012∗∗∗

(0.00000) (0.001)

F-Stat 14784.3 16782.2Observations 2,022,463 2,305,220R2 0.595 0.534Adjusted R2 0.010 −0.075

∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01Note: This table uses the same IV approach as Table 11 to show the effect of listing characteristics on theprice and rating of a guest’s next booking. All regressions control for market x check-in-date x how farin advance the search happened, the search date, the hour of the search interacted with the timezone thesearcher was searching from and any search filters the searcher used.

40

ReferencesArmstrong, M. (2006). Competition in two-sided markets. The RAND Journal of Economics,

37(3):668–691.Bardey, D., Cremer, H., and Lozachmeur, J.-M. (2009). Competition in two-sided markets

with common network externalities.Barnett, M. L. (2007). Tarred and untarred by the same brush: Exploring interdependence

in the volatility of stock returns. Corporate Reputation Review, 10(1):3–21.Cabral, L. and Hortacsu, A. (2010). The dynamics of seller reputation: Evidence from ebay.

The Journal of Industrial Economics, 58(1):54–78.Dellarocas, C. (2003). The digitization of word of mouth: Promise and challenges of online

feedback mechanisms. Management Science, 49(10):1407–1424.Evans, D. S. (2003). The antitrust economics of multi-sided platform markets. Yale Journal

on Regulation, 20:325.Fradkin, A., Grewal, E., Holtz, D., and Pearson, M. (2015). Bias and reciprocity in online

reviews: Evidence from field experiments on airbnb. In Proceedings of the Sixteenth ACMConference on Economics and Computation, pages 641–641. ACM.

Hagiu, A. (2007). Merchant or two-sided platform? Review of Network Economics, 6(2).King, A., Lenox, M. J., and Barnett, M. L. (2002). Strategic responses to the reputation

commons problem. Organizations, Policy and the Natural Environment: Institutional andStrategic Perspectives, pages 393–406.

Landon, S. and Smith, C. E. (1998). Quality expectations, reputation, and price. SouthernEconomic Journal, pages 628–647.

Luca, M. (2017). Designing online marketplaces: Trust and reputation mechanisms. Inno-vation Policy and the Economy, 17(1):77–93.

Mayzlin, D., Dover, Y., and Chevalier, J. (2014). Promotional reviews: An empirical investi-gation of online review manipulation. The American Economic Review, 104(8):2421–2455.

Morris, C. N. (1983). Parametric empirical Bayes inference: Theory and applications. Jour-nal of the American Statistical Association, 78(381):47–55.

Nosko, C. and Tadelis, S. (2015). The limits of reputation in platform markets: An empiricalanalysis and field experiment.

Rochet, J.-C. and Tirole, J. (2003). Platform competition in two-sided markets. Journal ofthe European Economic Association, 1(4):990–1029.

Rochet, J.-C. and Tirole, J. (2006). Two-sided markets: a progress report. The RANDJournal of Economics, 37(3):645–667.

Rysman, M. (2009). The economics of two-sided markets. The Journal of Economic Per-spectives, 23(3):125–143.

Shapiro, C. (1982). Consumer information, product quality, and seller reputation. The BellJournal of Economics, pages 20–35.

Ursu, R. M. (2016). The power of rankings: Quantifying the effect of rankings on onlineconsumer search and purchase decisions. Browser Download This Paper.

Weyl, E. G. (2010). A price theory of multi-sided platforms. The American EconomicReview, 100(4):1642–1672.

Winfree, J. A. and McCluskey, J. J. (2005). Collective reputation and quality. AmericanJournal of Agricultural Economics, 87(1):206–213.

41